Tips using Links

Fine-tuning

Models in ChainerCV support the argument pretrained_model to load pretrained weights. This functionality is limited in the case when fine-tuning pretrained weights. In that circumstance, the layers specific to the classes of the original dataset may need to be randomly initialized. In this section, we give a procedure to cope with this problem.

Copying a subset of weights in a chain can be done in few lines of code. Here is a block of code that does this.

# src is a model with pretrained weights
# dst is a model randomly initialized
# ignore_names is the name of parameters to skip
# For the case of VGG16, this should be ['/fc7/W', '/fc7/b']
ignore_names = []
src_params = {p[0]: p[1] for p in src.namedparams()}
for dst_named_param in dst.namedparams():
    name = dst_named_param[0]
    if name not in ignore_names:
        dst_named_param[1].array[:] = src_params[name].array[:]

Fine-tuning to a dataset with a different number of classes

When the number of classes of the target dataset is different from the source dataset during fine-tuning, the names of the weights to skip can be found automatically with the following method.

def get_shape_mismatch_names(src, dst):
    # all parameters are assumed to be initialized
    mismatch_names = []
    src_params = {p[0]: p[1] for p in src.namedparams()}
    for dst_named_param in dst.namedparams():
        name = dst_named_param[0]
        dst_param = dst_named_param[1]
        src_param = src_params[name]
        if src_param.shape != dst_param.shape:
            mismatch_names.append(name)
    return mismatch_names

Finally, this is a complete example using SSD300.

from chainercv.links import SSD300
import numpy as np

src = SSD300(pretrained_model='voc0712')
# the number of classes in VOC is different from 50
dst = SSD300(n_fg_class=50)
# initialized weights
dst(np.zeros((1, 3, dst.insize, dst.insize), dtype=np.float32))

# the method described above
ignore_names = get_shape_mismatch_names(src, dst)
src_params = {p[0]: p[1] for p in src.namedparams()}
for dst_named_param in dst.namedparams():
    name = dst_named_param[0]
    if name not in ignore_names:
        dst_named_param[1].array[:] = src_params[name].array[:]

# check that weights are transfered
np.testing.assert_equal(dst.extractor.conv1_1.W.data,
                        src.extractor.conv1_1.W.data)
# the names of the weights that are skipped
print(ignore_names)